What are the two core components of the DiffShape model, and how do they contribute to its overall objective?

The two core components are the self-supervised diffusion-based learning module and the contrastive language-shapelet learning module. The diffusion module generates shapelets from unlabeled subsequences using conditions derived from highly similar real subsequences, improving robustness when labels are scarce. The contrastive module aligns shapelet representations with language embeddings obtained from textual descriptions, strengthening the discriminative power of shapelets for classification.

DiffShape combines a self-supervised diffusion-based learning module with a contrastive language-shapelet learning module. In the diffusion module, fixed-size sliding windows turn each raw time series into subsequences, a convolutional layer extracts initial shapelets S0, and cosine similarity selects the most similar real subsequences Sr. Sr is used as a condition to a one-dimensional U-Net, so the generated shapelets are guided by real patterns from unlabeled data; the training loss is a reweighted noise-estimation loss. This lets the model learn shapelet distributions from large amounts of unlabeled data, reducing dependence on labels. In the contrastive module, textual descriptions are generated using ground-truth labels and classifier-derived pseudo-labels. A fixed pretrained language encoder maps these descriptions into language embeddings, while a shapelet encoder maps both original and generated shapelets into shapelet representations. Contrastive learning aligns the shapelet representations with the language embeddings, making the shapelets more discriminative for classification. The aligned representations are simultaneously used for downstream classification, so the two components jointly address semi-supervised time series classification with scarce labels and improve interpretability.